Out-of-Distribution Generalization and Detection With Limited Source Data

Out-of-distribution (OOD) generalization and detection have received significant attention in recent years, focusing primarily on addressing covariate shifts and category shifts in modern machine learning problems, respectively. However, most existing methods still rely on utilizing large amounts of source data to enhance the model generalization capabilities while effectively detecting unknown categories. To alleviate this constraint, this article proposes a novel method that simultaneously addresses covariate shifts and category shifts in OOD generalization and detection with limited source data. First, we design a distribution estimation method that leverages the weights of the pretrained classifier and an auxiliary dataset to generate estimated in-distribution (ID) feature representations. Subsequently, OOD generalization and detection are achieved by solving a constrained optimization problem based on the feature representations of the estimated wild data, where the estimated wild data comprise a mixture of estimated ID, covariate-shifted OOD, and category-shifted OOD data. Extensive experiments demonstrate that the proposed method can achieve competitive results even when only limited source data are available, as evidenced by comparisons with baselines specializing in either OOD generalization or OOD detection.